With MateCat machines and humans are stronger working together

Take care of the creative and highly specialized parts of translation. MateCat uses Machine Learning to handle repetitive tasks automatically.

Back in July, I already wrote about AdaptiveMT, one of the two new major features of the newly released SDL Trados Studio 2017. Since then that feature has been refined more and works well in the language combinations it presently supports (EN> FR, IT, NL, DE, ES).

Of course, one thing you'll need to take into account when thinking about using this feature is that it requires an additional paid subscription to the SDL Language Cloud -- much like what Google Translate also requires with its cost to use its API (which in turns makes it possible to use that service within a translation environment tool).

There are many different machine translation engines you can connect to SDL Trados Studio (here are the available plugins; the connectors to Google Translate and the SDL-provided options are installed by default), but if you have any question about which MT engine SDL would prefer you to use, you can lay it to rest when this information comes up if you try to select anything but SDL's engines:

Just necessary legal wording, you might argue. Well, I'm not completely sure about that. I've written a number of times about the fact that Google does not use your data when you use its API (which you do when you use it in SDL Trados Studio or another translation environment). Microsoft, on the other hand, does use the data it has access to, as may other MT engines, but I get a little frustrated with the apparent disregard of Google's position by other machine translation vendors such as Slate Desktop or, in this case, SDL. True, there might be all kinds of reasons not to use Google Translate within a translation environment, but confidentiality happens not to be one of those. Intimating that it is smacks of scaremongering to me.

(I'll step off my little pedestal now -- as if I needed one in the first place, being a proud 2 m/6'7".)

. . . continued below . . .

Webinar with limited places!

How do I find translation variants and what do I do with them?

Same old, same old? Yes and no.

During this Webinar, you will learn how to handle translation variants in Transit and how you can use them to provide better quality.

To register, simply send an e-mail to transit@star-group.net 

The other big feature in SDL's new version of Trados Studio is something very interesting as well. It's called "UpLIFT".

At the Translating Europe conference mentioned above, I was asked in an interview about the most important changes in translation technology during the last couple of years. If you've been reading this Journal for some time, you already know my answer: the introduction of machine translation into our workflows, in particular through suggestions of fragments, and the redefinition of translation memories as repositories not only of complete segments that can (but most likely won't) be turned into perfect and fuzzy matches but also subsegments (=fragments).

Hands down these are the two innovations that have had (and will continue to have) the most profound impact on how many of us work (keyword: "conductor" in the introduction).

In the first implementation of SDL Trados Studio (2009), the subsegmenting feature that was introduced was through "AutoSuggest" dictionaries. These were (are) standalone databases that were generated on the basis of relatively large translation memories (at first at least 25,000 translation units and with later releases of at least 10,000 TUs), and they were used solely to fuel an auto-suggest process (thus the name) where the translator was displayed fragments from her TMs via the AutoSuggest dictionaries as she typed. The process worked (and still works) well, but it had several drawbacks: The harvested bounty of fragments was not nearly as large as it could and should have been, there was never any context displayed for the fragments, and, most importantly, the AutoSuggest dictionaries were not automatically updated with newly translated content. Instead, you had to run the process of generating the AutoSuggest dictionaries again and again if you were interested in current data.

With UpLIFT, this is all completely different. First, the amount of data that is squeezed out of the translation memories that you process to gain subsegment data (you'll need a minimum of 5,000 translation units for decent results) is astonishingly large. Second, as with AutoSuggest, the data is offered as auto-suggestions as you type, but at the same time it's also displayed in a separate Fragment Matches pane where you can also see examples of the context in which the term or the phrase originally appeared. And third -- and again, most importantly -- each translated segment you submit is processed for subsegments as well; if applicable, these are then shown for later matches.

It's really quite helpful during translation, and I'm almost certain that it will garner a lot of praise from users of SDL Trados Studio 2017.

It may garner even more praise once users realize that the UpLIFT fragments (along with termbase matches and MT fragments) can also be used to repair fuzzy match translations (set under Project Settings/Options> Language Pairs> All Language Pairs> Match Repair)

What does this mean?

It means that if you have the following segment in the TM:

"If you start your computer from installation media that you created when you downloaded the ISO, such as a DVD or USB flash drive, you won't be able to keep your apps, Windows settings, or personal files when you install Windows 8.1."

And then you encounter this fuzzy-match segment (see the italicized words that have been changed):

"If you start your machine from installation media that you created when you downloaded the ISO, such as a DVD or USB flash drive, you won't be able to keep your apps, Windows settings, or personal folders when you install Windows 8.1."

. . . the system would intend to turn that fuzzy match into a perfect match by exchanging the "offending" terms with the correct ones in the translation. It does this by first looking at the termbase and fragment repository, and then, if nothing is found, into an associated machine translation engine if that feature is enabled. If machine translation is involved, it will not try to generate a suggestion for the whole segment but only for the two terms that have to be changed.

The resulting match still has to be reviewed of course and, depending on your language combination and the morphological complexity of the target language, possibly edited. But I'm sure you can easily see the value of this.

SDL Trados Studio is not the only tool that offers this auto-repair feature. Déjà Vu has been offering this for a long time, and memoQ's "MatchPatch" feature does the same thing (albeit without the option of an MT-based repair). I have tested this exact segment in all three tools and they all corrected the fuzzy match (almost) correctly (they all struggled with the plural of "folders," so I had to change the German "Ordnern" to "Ordner").

This implies that both Déjà Vu and memoQ also have a subsegmenting or fragment feature, as has been the case for some time. While SDL Trados Studio was the first to come out with the AutoSuggest feature, the other tools quickly followed suit and introduced their own comparable features (Déjà Vu's is called "DeepMiner" and memoQ's is the "Muse"). I spent a bit of time comparing SDL's new UpLIFT feature with DeepMiner and the Muse with the same underlying data.

All of them continue to update their fragment resources as you translate. Neither Déjà Vu nor memoQ has a minimum size requirement for the underlying TM, and especially Déjà Vu has good recall (and fuzzy-match-fixing) abilities with just a very small amount of data. With a large amount of data to process, however, DeepMiner's recall ability is not particularly strong in comparison with the other tools.

Let's use the example sentence above (with no machine translation or termbase connected but a domain-specific TM and its fragments) and assume that this is the correct German translation:

Wenn Sie Ihren Computer von den Installationsmedien starten, die Sie beim Download der ISO erstellt haben, wie etwa eine DVD oder ein USB-Flashlaufwerk, ist es nicht möglich, bei der Installation von Windows 8.1 Ihre Apps, Windows-Einstellungen oder persönlichen Dateien beizubehalten.

Déjà Vu only gave me "USB-Flashlaufwerk" and "Einstellungen" as the only helpful suggestions, both of which were actually not fragments but whole segments in the translation memory.

memoQ, on the other hand, gave me mostly single terms as auto-suggestions from the Muse. Terms that were missing were "Download," "DVD," "persönliche," and "Windows 8.1" -- all of which were part of the suggestions from Trados Studio. Also, Trados Studio displayed every auto-suggested term with its context in the Fragment Matches pane.

One drawback of the Trados Studio system is that it's overly complex to set the system up (and to find where to set it up) and that it's not activated by default.

Here are some of the steps you have to do to train the fragment recall:

Clearly, however, Studio'sUpLIFT feature is a real winner.

Some of you might still be puzzled, though. You might remember that all the way back in the beginning of 2014 I wrote about a tool called "Lift" that was designed by then-PhD student and translator Kevin Flanagan, and later on I mentioned that Kevin was hired by SDL to integrate his system.

Here's what I wrote about it:

Why not use external data to evaluate each segment, regardless of whether there is any existing data in that particular TM, and come up with suggested matches? That's what Kevin's Lift product does.

After processing with a tokenizer and, if applicable, a lemmatizer, single words are sent to a combination of machine translation (for those single words only) and/or dictionaries. These dictionaries can be termbases, other locally installed dictionaries, or even online dictionaries; if there are few or no dictionaries for your language combination, a pivot language can be used. On the basis of that invisible process and its resulting feedback, internal proposals are made about which part of the target segment goes with which part of the source segment (as required by a new segment that contains a part of that segment). More internal processes further verify the likelihood of these proposals before a highly educated guess is presented to the translator.

(...)This process breaks up the mold of the data that we produce -- in the form of a TM -- and analyzes it with outside data to make it more effective. This means that we still use the -- presumably -- high-quality translation that came out of our own fingertips or was dictated into our microphones, but we get to analyze it at an even deeper level than has been possible so far.

Sound a little different? Indeed it is. It turns out that a modified version of the original Lift feature will be implemented in server-based Trados Studio TMs in early 2017, but UpLIFT for the desktop-based version does not rely on external data and is more agile that way (but also less accurate and relies on more underlying TM data). The server-based UpLIFT feature (if it indeed will be called that also) will really be an entirely different product that will not care about looking at internal TM data to analyze fragment matches but will use word alignment and external dictionaries to decide what goes with what. It should be good -- but for the time being the desktop-based UpLIFT works great as well.

ATA Conference in San Francisco: Win cool prizes at the Across booth!

The 57th Annual Conference of the ATA will be held in San Francisco. More than 1,800 participants from all over the world can look forward to interesting presentations and an accompanying exhibition. At booth 13, Janine and Lena of Across will welcome visitors and answer questions about the Across Translator Edition and crossMarket.

Register at atanet.org 

Since their debut in 2007, smartphones have continued to gobble up or render obsolete all sorts of devices and other things we used to carry around with us. Point-and-shoot cameras, video cameras, radios, GPS receivers, pedometers, watches, magazines, newspapers, books, audiobooks, credit cards, and even cash can all be replaced by an app on a smartphone.  

As early as 2009 I can remember having conversations with interpreters and interpreting equipment manufacturers about the possibility of using smartphones as simultaneous interpretation receivers.

I stake no claim to this idea, as I am certain that many other interpreters and equipment manufacturers and end-users the world over surely had the same idea after the advent of the smartphone. In the last five years at least two equipment manufacturers and a handful of startups have built and begun marketing some form of interpreting platform that includes the use of a smartphone as a receiver for delegates who need to listen to interpretation in order to participate in a meeting, whether on site or virtually.

Each of the offerings take a somewhat different technical approach. They often have a different target market and use case, but they all include attendees' smartphones as receivers. It would be naive to think that the bring-you-own-device (BYOD) trend so prevalent in companies and organizations all over the world wouldn't at some point spill over into the conference interpreting equipment market.

None of these new offerings has been widely adopted yet, but one thing is certain, many meeting organizers and attendees are looking for alternatives to the old model of renting the standard set of interpreting equipment usually offered by audiovisual companies. They are looking to reduce the cost of equipment rental and offer their clients something more in line with their 21^st century technological expectations.

This month, I take an introductory look at Linguali-IS, one of these new platforms. In future installments of the Tech-Savvy Interpreter I'll look at other offerings as well.

The Linguali Concept

The Linguali (pronounced Ling-GWALL-lee) Interpreting System (Linguali-IS) makes full use of the BYOD concept. It eliminates the need for a client to rent receivers and headsets, interpreter consoles, audio mixing equipment and FM transmitters or infrared radiators. Meeting participants use their own smart devices (phones, tablets, etc.) as receivers. A laptop computer and headset replace the interpreter console or desk. Instead of using radio frequency or infrared light to transmit the interpretation to delegates, Linguali uses the existing local Wi-Fi network or a dedicated router that establishes its own network to send and receive audio. I'll dispense with the technical details in this initial look at Linguali, but the upshot is that the system uses digital voice over Internet protocol (VoIP) technology to work its magic. It's also important to note that Linguali-IS is not a remote interpreting platform. It simply replaces all the hardware needed to provide simultaneous interpretation within a meeting room.

Their business model seeks to address concerns expressed by many clients over the years, namely, the overall cost of providing simultaneous interpretation, particularly for smaller meetings with limited budgets. Linguali focuses on reducing equipment costs, thereby alleviating the often strong pressure placed on interpreters to reduce their rates in order to remain competitive. Rather than provide and charge for the rental of the equipment up front, Linguali charges for its service on a per connection basis. That is, if you expect 15 people to listen to interpretation at an event and only five actually do, you only pay for the five connections, not the 15 you originally expected. This is where the savings begin to accrue. Clients frequently overestimate the number of receivers they will likely need at an event and consequently pay to have a lot of equipment shipped, which often ends up sitting unused throughout the event, just to be on the safe side. Linguali's pay-only-for-what-you-use approach helps rationalize expenses and show the client that they only pay for the service the actually receive. If this platform and business model take off, equipment companies will face new competition in an area where they had a virtual monopoly in the past.

In addition to use by meeting organizers and interpreting agencies, the Linguali-IS platform is simple enough for independent interpreters to offer it to their direct clients as well. It is an alternative to purchasing "bidule" or portable equipment that requires an initial investment of a couple thousand dollars. Even so, I still see a market niche for interpreters that want to purchase their own portable interpreting equipment and rent it to their direct clients. So, I doubt these new Wi-Fi systems will totally replace the existing portable equipment so commonly used today.

The platform is currently in beta. If you are interested, you can download and install the software on your computer and smart devices and test it out on your home or office wireless network. The demo version only lets you connect for five minutes, but that is ample time to try out the platform. If you decide to use the platform professionally, Linguali currently charges 20 euro a month with an unlimited number of meetings and connections. Once out of beta, the platform will charge a fee of two euro per connection during an interpreted event.

Yeah, But...

Whenever I've discussed a Wi-Fi-based BYOD solution for simultaneous interpreting with colleagues, certain questions have always emerged. I'll try to address the main concerns here. All are valid questions, but the challenges they pose are not insurmountable.

Will participants want to download the app on their smart device?

Downloading an app is simple enough and I think it is certainly preferable to leaving a passport or driver's license with a technician in exchange for a receiver and headset during a conference. 

What if participants don't have headphones or earbuds with them?

Most smartphone users carry a set of earbuds with them almost everywhere they go now. (Take a ride on a metro at rush hour in any large metropolitan area in the world if you want proof.) If participants don't have a pair with them, adequate inexpensive pairs can be provided a little or no cost.

Won't battery life be an issue?

Battery life is probably the trickiest question on the list. It can be a problem if attendees come with phones that are not fully charged, or the meeting is going to last more than a few hours. Plugging in your smart device is a simple solution, but it is not always an option, since many meeting rooms don't have many electrical outlets. Additional batteries may also be necessary. This is why I think the BYOD solution is best suited for meetings that last 4 hours or less. Actual experience may prove me wrong on this, but only time will tell.

Will the platform be simple enough for participants to use without a lot of handholding?

If the beta version of Linguali-IS is any indication, yes. In fact, the current interfaces are a little too simple for my taste. The smartphone app interface for participants is extremely simple. Open the app and it automatically connects as long as it is on the same Wi-Fi network as the computer being used as an interpreter console. Here's a screenshot of what a participant sees. It has a simple channel selector, volume control and microphone mute button. That's it.

If you are wondering what the microphone mute button is for, well, the smart device can also serve as a microphone that can pick up the delegates voice and transmit it to the interpreter. Smartphone microphones are not always the best quality, but that direct sound feed beats interpreting ambient sound that you pick up with just the naked ear.

The interpreter interface is also bare-bones simple and could benefit from the inclusion of some additional features without making it too complex. The good news is that this is easy enough to do since Linguali-IS is software that can be rewritten and tweaked to introduce improvements as they continue to beta test. Here's a screenshot of the current interpreter interface.

How many participants can connect to the Wi-Fi network before it becomes congested?

Linguali recommends that no more than 20 participants connect to the Wi-Fi network used for simultaneous interpretation. Congestion is another reason why having a separate network only for the interpretation is a good idea. The audio feed shouldn't have to compete with social media posts and email checking that goes on most conference networks. This limit is another reason why I think Linguali-IS in its current form is best suited for smaller meetings, not big, multi-day conferences with thousands of participants.  

My Take

There is definitely a niche for this kind of tech. However, it is so new and innovative that many potential clients may be hesitant to become early adopters. Linguali-IS has the potential to expand access to simultaneous interpreting by reducing the associated equipment costs and allowing interpreters to still charge a fair professional fee for their services.   

Many conference interpreters continue to think about interpreting jobs in terms of days. But the reality is that there are more and more meetings being organized with simultaneous interpretation that last 3 hours or less. Their budgets are smaller, and organizers are keen to offer simultaneous interpretation if they can find a cost-effective way to do it. VoIP audio over a local Wi-Fi network combined with attendees bringing their own devices may prove to be the solution. If so, many equipment providers and event organizers will have to adjust in order to stay competitive in this market. Tech-savvy event organizers and entrepreneurial interpreters would do well to start experimenting with this technology now.

Do you have a question about a specific technology? Or would you like to learn more about a specific interpreting platform, interpreter console or supporting technology? Send us an email at inquiry@interpretamerica.com.

Join GALA and InterpretAmerica for think! Interpreting 2017 in Amsterdam!

The 4^th installment of the groundbreaking GALA/InterpretAmerica collaboration think! Interpreting will take place at the 2017 Language of Business Conference in Amsterdam from March 26-29, 2017.

Look for the preliminary program in late November. Then be sure to join us in 2017! 

I was also asked in Brussels to predict one fundamental change I thought we'd experience in the world of translation due to technology. Again, this was not something I had to think very hard about. To me (as many of you know) it's clear that technology will force us to change our pricing structure. To stay with our metaphor in the introduction, it is theoretically possible to count the number of notes played by one violist and pay him accordingly. (Sound ridiculous? That's exactly how we are paid!) But when you're the conductor of the whole orchestra, those units of measurement, as ridiculous as they might have been in the first place, don't work anymore. That's when you have to start being paid in a different manner, whether project-based or time-based.

Lilt recently introduced a new version and pricing scheme as well as a number of new features, including one that I particularly like called the "Quote Builder." The Quote Builder is a tool that combines a number of metrics to inform you, the translator, what you should be charging for a project and/or how much value you produce.

How does it do it? It looks at a number of metrics, most importantly your average translation speed in earlier projects. It then takes that number and associates it with the number of words in the current project, perfect/fuzzy match rates percentages (if so desired), and your per-word price, and calculates your hourly rate on that basis.

What is that good for?

For one thing, it's a good idea for you to know how much money per hour you make. This is good for your own information and also to understand whether you need to adjust your prices to be at an acceptable level. That is, if you still charge by the word. You see, I think Quote Builder's most important function is as a transitional tool of sorts to help us understand what we need to charge by the hour (or project). Many of you might have a very clear idea about what you need to charge per hour to be profitable (judging from the questions at Translating Europe and elsewhere you're in the minority, though), but it clearly needs to be in some kind of relationship with what your clients have been willing to pay so far. This is a fabulous tool that allows you to show clients why it's OK for you to charge $80 or $60 or $40 per hour (or $200 for that matter) -- you are essentially using your own track record that you've built up in Lilt to give them an idea of your productivity.

I am sure that we'll see tools like this cropping up left and right, and I'm proud of Lilt for setting the bar for us very early in the game. 

In the last Tool Box Journal I added a comment to Riccardo's excellent article about Xbench to the effect that Xbench's integration with Memsource is not available for the translator editions. Silly me! It's been available for quite a while as you can see here.

A few editions ago, I wrote about BaacS from Belarus, a free web-based invoice and accounting software for translators. I had mentioned that Eugene Kuchynski, the developer behind it, was also working on a desktop version for Windows that he would be releasing later this year. It ended up being released earlier rather than later. It costs $33 until the end of the year and will then be a modest $40. What I liked particularly about this tool was the possibility of migrating data from Translation Office 3000 in case you have used that tool before and are looking for a change.

I got a few responses regarding the article about the eclectic "Spike" feature that MS Word offers (and hides very well). Paul Makinen sent me a link to a posting that seems to contain a truly exhaustive list of things you can do with that. Thanks!

AdaptiveMT will be your own machine translation engine that adapts to suit your style, content and terminology.

Each time you make a change to MT output your engine learns from it and applies it in real-time to the next sentences. 

As a subscriber to the Premium version of this journal you have access to an archive of Premium journals going back to 2007.

You can access the archive right here. This month the user name is toolbox and the password is hearthemusic.

New user names and passwords will be announced in future journals.

If you would like to promote this journal by placing a link on your website, I will in turn mention your website in a future edition of the Tool Box Journal. Just paste the code you find here into the HTML code of your webpage, and the little icon that is displayed on that page with a link to my website will be displayed.

Here are some folks who have recently linked to the Journal:

blog.gauchatranslations.com

ecpdwebinars.co.uk

If you are subscribed to this journal with more than one email address, it would be great if you could unsubscribe redundant addresses through the links Constant Contact offers below.

Should you be interested in reprinting one of the articles in this journal for promotional purposes, please contact me for information about pricing.

© 2016 International Writers' Group